Satellite image target detection system based on deep learning

Abstract

The application relates to the technical field of satellite image target detection, and discloses a satellite image target detection system based on deep learning. The system collects vehicle satellite image data Clsj and airplane satellite image data Fjsj through a data acquisition module, carries out image enhancement on the satellite images through a data preprocessing module, obtains airplane characteristic values Fjtz and vehicle characteristic values Cltz through feature extraction based on the data transmitted by the data preprocessing module through a deep convolutional neural network, carries out target detection algorithm optimization through a target detection algorithm model optimization module, and outputs the detected airplane and vehicle satellite images based on the optimized algorithm model through a visualization module. The satellite images of vehicles, airplanes and the like are detected through the optimized algorithm model, and the accuracy of target image detection is improved.

Description

Technical Field

[0001] This invention relates to the field of satellite image target detection technology, specifically a satellite image target detection system based on deep learning. Background Technology

[0002] Remote sensing is a long-range, non-contact detection technology that uses infrared, microwave, and other detection techniques to identify targets through photography, scanning, and information sensing. Satellite remote sensing is one such technology, providing continuous observation of large areas and applicable to various downstream remote sensing applications such as traffic management, marine monitoring, smart cities, and aviation. However, due to the massive amount of satellite image data and the scarcity of valuable information (i.e., low data value density), processing is extremely difficult. Therefore, it is essential to utilize intelligent auxiliary tools to help people extract information from satellite images and improve efficiency.

[0003] With the development of science and technology and the continuous progress of computer software and hardware, artificial neural networks have emerged. Artificial neural networks are mathematical models that scientists build on computers that can learn and imitate human thinking. Deep learning, a branch of machine learning, has also emerged. Deep learning allows computers to learn the meaning and patterns behind data, that is, it gives computers the ability to think for themselves. This powerful ability has expanded the field of deep learning to almost all machine learning-related fields. The field of object detection has also undergone a qualitative change after the addition of deep learning.

[0004] However, achieving accurate and powerful moving target detection and tracking in satellite imagery remains challenging. Satellite imagery data itself has extremely high resolution, while targets to be detected, such as vehicles and aircraft, are generally only a dozen pixels in the image, resulting in low recognition and low detection accuracy. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a satellite image target detection system based on deep learning. The system comprises a data acquisition module that acquires vehicle satellite image data (Clsj) and aircraft satellite image data (Fjsj); a data preprocessing module that enhances the satellite images; a feature extraction module that extracts features from the data transmitted by the data preprocessing module using a deep convolutional neural network to obtain aircraft feature values ​​(Fjtz) and vehicle feature values ​​(Cltz); a target detection algorithm model optimization module that optimizes the target detection algorithm; and a visualization module that outputs the detected aircraft and vehicle satellite images based on the optimized algorithm model. By using the optimized algorithm model to detect satellite images of vehicles, aircraft, and other targets, this system improves the accuracy of target image detection and solves the aforementioned problems.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention provides the following technical solution: a satellite image target detection system based on deep learning, characterized in that it includes a data acquisition module, a data preprocessing module, a feature extraction module, a target detection algorithm model optimization module, and a visualization module;

[0009] The data acquisition module is connected to the data preprocessing module via a network;

[0010] The data preprocessing module is connected to the feature extraction module via a network;

[0011] The feature extraction module is connected to the target detection algorithm model optimization module via a network;

[0012] The target detection algorithm model optimization module is connected to the visualization module via a network;

[0013] The data acquisition module is used to acquire vehicle satellite image data Clsj and aircraft satellite image data Fjsj, and transmit the acquired data to the data preprocessing module via the network.

[0014] The data preprocessing module receives vehicle satellite image data Clsj and aircraft satellite image data Fjsj transmitted by the data acquisition module, performs image enhancement on the above data to obtain enhanced vehicle satellite image Zqcl and enhanced aircraft satellite image Zqfj, and the data preprocessing module transmits the above processed data to the feature extraction module.

[0015] The feature extraction module extracts features based on the data transmitted by the data preprocessing module through a deep convolutional neural network to obtain aircraft feature value Fjtz and vehicle feature value Cltz, and transmits the calculated feature values ​​to the target detection algorithm model optimization module.

[0016] The target detection algorithm model optimization module optimizes the target detection algorithm model based on the feature values ​​transmitted by the feature extraction module and in combination with the optimization algorithm.

[0017] The visualization module outputs satellite detection images based on the optimized target detection algorithm model.

[0018] Preferably, the data acquisition module compiles the acquired vehicle satellite images into a vehicle image dataset, which is {Clsj} n-9 Clsj n-8 Clsj n-7 ,...,Clsj n};

[0019] The data acquisition module compiles the acquired aircraft satellite image data into an aircraft image dataset, which is {Fjsj}. n-9 Fjsj n-8 Fjsj n-7 , ..., Fjsj n}

[0020] Preferably, the data preprocessing module performs image enhancement on the vehicle satellite image data Clsj, and the image enhancement formula is as follows:

[0021] Zqcl=output_pixel=input_pixel+delta_b

[0022] In the formula, Zqcl represents the enhanced vehicle satellite image, output_pixel represents the enhanced pixel value, input_pixel represents the original image pixel value, and delta_b represents the brightness increment.

[0023] Preferably, the data preprocessing module performs image enhancement on the aircraft satellite image data Fjsj, and the image enhancement formula is as follows:

[0024] Zqfj=output_pixel=contrast_factor*(input_pixel-mean_pixel)

[0025] In the formula, Zqfj represents the enhanced aircraft satellite image, output_pixel represents the enhanced pixel value, contrast_factor represents the contrast factor, input_pixel represents the original image pixel value, and mean_pixel represents the pixel mean.

[0026] Preferably, the feature extraction module extracts features based on the data transmitted by the data preprocessing module using a deep convolutional neural network to obtain aircraft feature values ​​Fjtz and vehicle feature values ​​Cltz. The calculation method is as follows:

[0027] S1. Input images: Input images of airplanes and vehicles and feed them into the convolutional neural network model.

[0028] S2, Convolution and Activation Functions: The image will pass through multiple convolutional layers and activation functions. In each convolutional layer, the image features will be convolved by different convolutional kernels and processed by activation functions.

[0029] S3. Pooling layers are usually added after convolutional layers to downsample the feature maps;

[0030] S4. The feature maps that have undergone convolution and pooling are integrated and abstracted through multiple fully connected layers to obtain the final feature representation.

[0031] S5. The output of the last layer can be represented as the feature vectors of the aircraft and the vehicle.

[0032] Preferably, the target detection algorithm model optimization module is used to optimize the target detection algorithm, and the optimization method is as follows:

[0033] The R-CNN algorithm extracts candidate regions for each image using Selective Search, and then each region is input into the CNN as a sample to extract features. The model structure introduces RoL Pooling for filtering to solve the problem of duplicate feature extraction. After further improvement, RPN (Region Proposal Network, Anchor-Based two-stage algorithm for extracting candidate boxes) is used instead of Selective Search.

[0034] Preferably, the target detection algorithm model optimization module uses the YOLO algorithm to optimize the algorithm model.

[0035] Preferably, the target detection algorithm model optimization module compares the Anchor-Based two-stage algorithm and the Anchor-Based single-stage algorithm.

[0036] Preferably, the target detection algorithm model optimization module optimizes the PP-YOLO model.

[0037] Preferably, the visualization module outputs satellite images of detected aircraft and vehicles based on the optimized PP-YOLO algorithm model.

[0038] Compared with existing technologies, this invention provides a satellite image target detection system based on deep learning, which has the following advantages:

[0039] This invention acquires satellite images of vehicles and aircraft through a data acquisition module. A data preprocessing module enhances these satellite images to obtain an enhanced vehicle satellite image Zqcl and an enhanced aircraft satellite image Zqfj. A feature extraction module extracts features based on the data transmitted from the data preprocessing module using a deep convolutional neural network to obtain aircraft feature values ​​Fjtz and vehicle feature values ​​Cltz. A target detection algorithm model optimization module optimizes the target detection algorithm. The optimized algorithm model is used to detect satellite images of vehicles and aircraft, avoiding the problems of low image recognition and difficulty in extracting feature values ​​during training when performing target image detection, thus improving the accuracy of target image detection. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the system flow of the present invention; Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] To address the problem of low target image recognition accuracy in current satellite image target detection methods, a deep learning-based satellite image target detection system is proposed. Please refer to [link to relevant documentation]. Figure 1 The system specifically includes a data acquisition module, a data preprocessing module, a feature extraction module, a target detection algorithm model optimization module, and a visualization module;

[0043] The data acquisition module compiles the acquired vehicle satellite images into a vehicle image dataset, which is {Clsj}. n-9 Clsj n-8 Clsj n-7 ,...,Clsj n};

[0044] The data acquisition module compiles the acquired aircraft satellite image data into an aircraft image dataset, which is {Fjsj}. n-9 Fjsj n-8 Fjsj n-7 , ..., Fjsj n};

[0045] Because satellite images generally suffer from insufficient clarity, image enhancement is required for vehicle satellite image data Clsj and aircraft satellite image data Fjsj through a data preprocessing module. The vehicle image enhancement formula is as follows:

[0046] Zqcl=output_pixel=input_pixel+delta_b

[0047] By enhancing satellite imagery, images can be made clearer and more contrasting, thereby increasing the details and information content in the image. This helps users obtain geographic information of target areas more accurately, promoting more in-depth geographic analysis and research. In the formula, Zqcl represents the enhanced vehicle satellite image, output_pixel represents the enhanced pixel value, input_pixel represents the original image pixel value, and delta_b represents the brightness enhancement. By enhancing the brightness, the brightness of the image is magnified to achieve the purpose of enhancing image clarity. The significance of enhancing satellite imagery lies in improving the reliability and applicability of data, and promoting the application and development of remote sensing technology in various fields.

[0048] The formula for aircraft image enhancement is as follows:

[0049] Zqfj=output_pixel=contrast_factor*(input_pixel-mean_pixel)

[0050] Enhanced satellite imagery can reduce noise and distortion, improving image quality and accuracy. This is crucial for professional work in fields such as remote sensing and geographic information systems, helping to ensure data reliability and availability. In the formula, Zqfj represents the enhanced aircraft satellite image, output_pixel represents the enhanced pixel value, contrast_factor represents the contrast factor, input_pixel represents the original image pixel value, and mean_pixel represents the pixel mean. By setting the contrast factor, the contrast of the aircraft satellite image is improved, thereby enhancing image clarity.

[0051] The feature extraction module, based on the data transmitted from the data preprocessing module, performs feature extraction using a deep convolutional neural network to obtain aircraft feature values ​​Fjtz and vehicle feature values ​​Cltz. The calculation method is as follows:

[0052] S1. Input images: Input images of airplanes and vehicles and feed them into the convolutional neural network model.

[0053] S2, Convolution and Activation Functions: The image will pass through multiple convolutional layers and activation functions. In each convolutional layer, the image features will be convolved by different convolutional kernels and processed by activation functions.

[0054] S3. Pooling layers are usually added after convolutional layers to downsample the feature maps;

[0055] S4. The feature maps that have undergone convolution and pooling are integrated and abstracted through multiple fully connected layers to obtain the final feature representation.

[0056] S5. The output of the last layer can be represented as the feature vector of the aircraft and the vehicle.

[0057] The object detection algorithm model optimization module is used to optimize the object detection algorithm. The optimization method is as follows:

[0058] The R-CNN algorithm extracts candidate regions for each image using Selective Search, and then each region is input into the CNN as a sample to extract features. Roll Pooling is introduced into the model structure to filter and solve the problem of duplicate feature extraction. After further improvement, RPN (Region Proposal Network, Anchor-Based two-stage algorithm for extracting candidate boxes) is used instead of Selective Search, thereby improving the efficiency of candidate box extraction.

[0059] The target detection algorithm model optimization module uses the YOLO algorithm to optimize the algorithm model;

[0060] A comparison of the object detection algorithm model optimization module between the Anchor-Based two-stage algorithm and the Anchor-Based single-stage algorithm is shown below:

[0061] Compared to the anchor-based two-stage algorithm, the anchor-based single-stage algorithm is significantly more complex. The two-stage algorithm requires a learning network (RPN) for candidate bounding boxes and a learning network (CNN) for object regression and classification, resulting in a much more complex network structure. The anchor-based single-stage algorithm, however, eliminates the RNP component, directly regressing and classifying the final bounding box and category from the anchors. Despite this, the anchor-based two-stage algorithm achieves superior recognition accuracy compared to the anchor-based single-stage algorithm, resulting in a recognition speed unmatched by other algorithms. The YOLO series of algorithms, due to their small model size, ease of deployment, and high recognition speed while maintaining satisfactory accuracy, are cost-effective choices and have been widely applied in industries such as manufacturing, transportation, and facial recognition.

[0062] The object detection algorithm model optimization module optimizes the PP-YOLO model, and the optimization scheme is as follows:

[0063] The PP-YOLO model has been optimized in six main aspects: data augmentation, data sampling / positive and negative sample sampling, network structure, learning strategy, training strategy, and post-processing, as detailed below:

[0064] Image Mixup: The idea is to align and stack ground truths of different categories in a certain proportion before training. This can improve the generalization ability of convolutional networks to spatial perturbations, increase the proportion of positive samples to a certain extent, and improve the recall rate.

[0065] Label Smooth: In classification problems, the traditional one-hot encoded label vector yhot is 1 if i is the target and 0 otherwise. This can lead to overfitting and poor generalization ability because the training cannot cover all cases. Label Smooth replaces it with a new perturbed label vector, thereby softening the label and improving accuracy.

[0066] Synchronized Batch Norm: In multi-GPU training, Batch Norm typically calculates the mean and variance independently for each GPU. However, Synchronized Batch Norm calculates the mean and variance synchronously for each GPU in a multi-GPU setup. Even when training with multiple GPUs, it's treated as if training with a single GPU. This results in more reasonable mean and variance calculations, leading to better convergence speed and performance.

[0067] ResNet-D: In ResNet, when using a 1×1 Conv, sliding two squares at a time causes a loss of 3 / 4 of the information in each downsampling stage. ResNet-D adjusts the sliding position to slide on a 3×3 Conv, ensuring that more information is retained in each downsampling, thereby improving recognition accuracy.

[0068] Deformable Convolution (DCN) learns the kernel shape by learning the convolution weights, which are typically a square shape. However, this square shape is not suitable for detecting all types of objects. Deformable Convolution aims to learn the kernel shape simultaneously with the weights. It learns the kernel shape by using an offset to determine the direction and distance of the convolution offset. This optimization results in a slight improvement in accuracy and a significant increase in prediction speed.

[0069] Drop Block: First, let's mention DropOut. DropOut is a commonly used training method in deep learning to reduce network overfitting and improve network generalization ability. Applying this idea to object detection results in Drop Block. Drop Block optimizes the approach because in object detection, learning the shape of a convolution is essentially learning a connected component. DropOut discards random discrete points, which is not suitable for this situation. Therefore, Drop Block computationally drops an entire region randomly instead of random discrete points on the convolution, making this method more suitable for object detection tasks that "detect connected components".

[0070] Exponential Moving Average: This is a method in value investing that measures price trends by calculating the exponential moving average of historical fluctuations. Used in model training, it allows for the calculation of historical moving averages on parameters, resulting in a smoother and more stable training process. The formula is as follows:

[0071] EMA0=0

[0072] EMA t =decay*EMA t-1 +(1-decay)*θ t

[0073] Larger Batch Size: By increasing the batch size parameter, the number of samples calculated in each iteration during training increases, thereby making the model training more stable and significantly improving accuracy without affecting the prediction speed.

[0074] IoU loss: When calculating the loss, the negative value of IoU (Interaction over Union, which evaluates the quality of the predicted bounding box) is added to the calculation. This way, while optimizing the loss to reduce it, the IoU is actually increased, ultimately achieving the optimization effect. The formula is as follows:

[0075] IoU loss=(1-IoU*IoU)*loss_weight

[0076] It can improve prediction accuracy without changing the prediction speed.

[0077] IoU Aware: In NMS, the specific formula for calculating the score is as follows:

[0078] score=objectness*classification

[0079] This means that the predicted box (PB) needs to contain an object that belongs to the corresponding category. This can lead to situations where the PB has a high IoU with the ground truth bounding box, but is filtered out in Non-Maximum Suppression (NMS) due to a low score. The IoU-Aware approach incorporates localization accuracy into the score calculation. Specifically, this involves adding a channel to learn the IoU, increasing the number of output channels from B*(5+C) to B*(5+C). During prediction, the predicted value of the objectness channel is also updated as follows:

[0080] S det =p i α *IoU i 1-α

[0081] Because this is an optimization technique used in the prediction process, the prediction speed is reduced, but the prediction accuracy is improved.

[0082] Grid Sensitive: YOLOv3 uses sigmoid activation when calculating the center point of the ground truth bounding box. If the center point falls exactly on the edge of the grid, the model will learn towards positive or negative infinity, leading to overfitting. Therefore, YOLOv4 introduced Grid Sensitive, adding an offset and scaling factor when calculating the center coordinates, as shown in the following formula:

[0083] Depend on

[0084] become

[0085] This is an optimization strategy used during training, so the prediction speed remains unchanged while the accuracy is slightly improved.

[0086] Matrix NMS: NMS works by first calculating the scores of a group of closely spaced boundary pairs (PBs) of the same category. The image with the highest score is then compared to the remaining images for an Intersection over Union (IoU). PBs with an IoU greater than a certain threshold are considered similar boxes and are eliminated. This process is repeated for the entire group of images, retaining only dissimilar PBs, i.e., valid PBs. However, when PBs are close together, those with lower scores are directly filtered out. This means that if two objects are close and belong to the same category, the PB with the lower score will be filtered out, reducing recognition accuracy. Therefore, Soft NMS was developed. Instead of direct suppression, it introduces a penalty coefficient, as shown in the formula below:

[0087] Depend on

[0088] become

[0089] This optimization improved the model's ability to identify nearby targets. However, Soft NMS is computationally much more expensive than before. Matrix NMS addresses this by using parallel computing on top of Soft NMS to optimize computation speed. Adjustments were made during the computation process, but due to its complexity, they will not be detailed here. This optimization occurred during training and therefore did not affect prediction speed, but improved prediction accuracy.

[0090] Coord Conv: Normally, convolution only perceives the receptive field corresponding to the convolution kernel. That is, if a dog appears in an image, it's a dog whether it's in the top left corner or the bottom right corner; the convolution doesn't care about its position in the image, only the receptive field of the convolution kernel. Therefore, if the dog is at the edge of the image and not fully visible, it cannot be accurately identified. Coord Conv adds two channels, inputting x and y coordinate information respectively, allowing the convolution to obtain the global coordinates of the target, thus facilitating the extraction of target features. Although this increases computational cost, it has a small impact on prediction speed and significantly improves accuracy.

[0091] SPP (Spatial Pyramid Pooling) is a feature extraction method that uses pooling windows of different scales. By extracting multiple features and then concatenating them, it achieves both an overall and a local perception, thus optimizing feature extraction. In PP-YOLO, SPP is only added during the backbone network's feature map output, resulting in only a minor impact on prediction speed while improving accuracy.

[0092] SSLD: Simple Semi-supervised Label Distillation. Here, we need to mention the concept of distillation. Distillation uses the classification probabilities learned by the large model as labels to "guide" the learning of the small model. This increases the model's accuracy without affecting the network structure of the small model. SSLD is an excellent knowledge distillation model released on PaddleClass. Adding it to the PP-YOLO network and performing distillation on PP-YOLO is an optimization technique during the learning process. It does not affect the prediction speed and significantly improves the model's prediction accuracy.

[0093] 2x Scheduler: 1x Scheduler represents the strategy used in training, which generally specifies the number of iterations and the learning rate, and uses different learning rates at different iteration stages. 2x Scheduler, as the name suggests, is based on 1x Scheduler, but adjusts all parameters to twice that of 1x Scheduler, which can significantly improve prediction accuracy without affecting speed.

[0094] According to the paper published in 2020, after continuous optimization, PPYOLO finally achieved an mAP of 45.9% on the COCO dataset and a single-card speed of 72.9 FPS on V100. Under the same conditions, it is significantly higher than YOLOv4's mAP of 43.5% and FPS of 62.0.

[0095] The visualization module outputs satellite images of detected aircraft and vehicles based on the optimized PP-YOLO algorithm model;

[0096] By using an optimized algorithm model to detect vehicles and aircraft in satellite images, the problems of low image recognition and difficulty in extracting feature values ​​during training are avoided, thus improving the accuracy of target image detection.

[0097] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A satellite image target detection system based on deep learning, characterized in that: It includes a data acquisition module, a data preprocessing module, a feature extraction module, a target detection algorithm model optimization module, and a visualization module; The data acquisition module is connected to the data preprocessing module via a network; The data preprocessing module is connected to the feature extraction module via a network; The feature extraction module is connected to the target detection algorithm model optimization module via a network; The target detection algorithm model optimization module is connected to the visualization module via a network; The data acquisition module is used to acquire vehicle satellite image data. With aircraft satellite image data And transmit the aforementioned vehicle satellite image data via the network With aircraft satellite image data The data is transmitted over the network to the data preprocessing module. The data preprocessing module receives vehicle satellite image data transmitted by the data acquisition module. With aircraft satellite image data Image enhancement is performed on the above data to obtain enhanced vehicle satellite images. With enhanced aircraft satellite imagery The data preprocessing module transmits the processed data to the feature extraction module. The feature extraction module extracts aircraft feature values ​​based on the data transmitted by the data preprocessing module using a deep convolutional neural network. Vehicle characteristic values and the above aircraft characteristic values Vehicle characteristic values Transmitted to the target detection algorithm model optimization module; The target detection algorithm model optimization module optimizes the target detection algorithm model based on the feature values ​​transmitted by the feature extraction module and in combination with the optimization algorithm. The visualization module outputs satellite detection images based on the optimized target detection algorithm model; The data preprocessing module processes vehicle satellite image data. Image enhancement is performed using the following formula: ， In the formula, This indicates enhanced satellite imagery of vehicles. This represents the enhanced pixel value. Represents the pixel values ​​of the original image. Indicates an increase in brightness; The data preprocessing module processes aircraft satellite image data. Image enhancement is performed using the following formula: ， In the formula, This indicates enhanced aircraft satellite imagery. This represents the enhanced pixel value. Represents the contrast factor. Represents the pixel values ​​of the original image. This represents the average pixel value.

2. The satellite image target detection system based on deep learning according to claim 1, characterized in that: The data acquisition module compiles the acquired vehicle satellite images into a vehicle image dataset, which is: ; The data acquisition module compiles the acquired aircraft satellite image data into an aircraft image dataset, which is: .

3. The satellite image target detection system based on deep learning according to claim 1, characterized in that: The feature extraction module extracts aircraft feature values ​​based on the data transmitted by the data preprocessing module using a deep convolutional neural network. Vehicle characteristic values The calculation method is as follows: S1. Input images: Input images of airplanes and vehicles and feed them into the convolutional neural network model. S2, Convolution and Activation Functions: The image will pass through multiple convolutional layers and activation functions. In each convolutional layer, the image features will be convolved by different convolutional kernels and processed by activation functions. S3. A pooling layer is added after the convolutional layer to downsample the feature map; S4. The feature maps that have undergone convolution and pooling are integrated and abstracted through multiple fully connected layers to obtain the final feature representation. S5. The output of the last layer can be represented as the feature vectors of the aircraft and vehicles.

4. The satellite image target detection system based on deep learning according to claim 3, characterized in that: The target detection algorithm model optimization module is used to optimize the target detection algorithm. The optimization method is as follows: pass The algorithm processes each image through... Extract candidate regions, and then input each region as a sample. Extract features from the model and introduce them into the model structure. This is used for filtering to solve the problem of duplicate feature extraction. Further improvements are then made using... Replaced .

5. A satellite image target detection system based on deep learning according to claim 4, characterized in that: The target detection algorithm model optimization module utilizes Algorithm optimization algorithm model.

6. The satellite image target detection system based on deep learning according to claim 5, characterized in that: Comparison of the target detection algorithm model optimization module Two-phase algorithm and Single-stage algorithm.

7. A satellite image target detection system based on deep learning according to claim 6, characterized in that: The target detection algorithm model optimization module is optimized. Model.

8. The satellite image target detection system based on deep learning according to claim 7, characterized in that: The visualization module is based on the optimized version. The algorithm model outputs satellite images of the detected aircraft and vehicles.

Images